Why AI Feels Like the Internet in 1997 | Benedict Evans on a16z

Agentic coding achieved sudden product-market fit

Evans identifies agentic coding as the first AI application that shifted from marginally useful to fundamentally transformative, creating a supply crunch where customers are pulling tools out of vendors' hands.

Developers build for themselves first

Following the historical pattern of early PCs—where the first application was building more computers—software developers are using LLMs to automate software development before other use cases mature.

Industry focus narrowed from broad to specific

After OpenAI's initial 'everything at once' approach, the field has consolidated around coding following Anthropic's focused strategy, though it remains unclear which use case will bridge the gap to daily consumer adoption.

$10 trillion spending ceiling creates hard constraints

Evans emphasizes that the industry faces absolute scarcity because there simply isn't $10 trillion annually available to spend on AI infrastructure, creating an unsustainable gap between demand and supply.

Pricing chaos mirrors mobile data crisis of 2009

Current market conditions—where users face either $20 monthly subscriptions or $10,000 surprise bills—exactly parallel the mobile data pricing breakdown before carriers implemented throttling and usage caps.

Capex overbuild threatens commodity pricing

With $1-2 trillion in capital expenditure coming online while models become 100x more efficient annually, infrastructure faces impending overcapacity that could eliminate pricing power for model providers.

Foundation models risk becoming commodity infrastructure

Lacking network effects and differentiation, models may follow the path of mobile operators and ISPs who built essential infrastructure but captured minimal value as the technology matured.

Historical precedent favors application layer value

Evans notes that chip companies, ISPs, and mobile operators collectively spent trillions on infrastructure yet saw flat stocks for decades, while operating systems like Windows and iOS captured value by enabling ecosystem lock-in.

Chatbots are not viable standalone products

Evans explicitly argues that chatbots and foundation models are not finished products but rather infrastructure components, comparing them to browsers that enabled value creation elsewhere rather than capturing it themselves.

Job market impacts remain unpredictable

While automation forces companies to reconsider why they hire junior engineers—to do tasks now automated or to train for future roles—Evans insists it is impossible to predict engineering team structures even three years out.

Enterprise adoption targets point solutions

Outside tech hubs, corporations deploy AI for specific back-office processes rather than general assistant use, such as commodities companies using LLMs exclusively for cash flow forecasting from irregular producer invoices.

Fundamental questions remain unanswered

Evans emphasizes that core uncertainties—including whether models can capture value up the stack, if there will be a single winner, and how to drive daily consumer usage—persist just as they did two years ago.